Suggestions not directly responsive to your question, but likely to be valuable nonetheless:

One basic strategy to consider, taking account of country-specific "norms" and what "The Authorities" might view as particularly suspicious, is using two or more devices for different roles.

For example, you might have a PC and printer in your home, but use a smart phone when covering street protests or meeting with sources. If so, when you are outside your home, you might use the new Tor for Android for urgent messages, to set up meetings, etc. (You might use a second smart phone in "monitor mode" for countersurveillance in the field, unless using two phones will look strange in your country.) When you are working at home, you might boot Tails from a DVD for further research on the internet, keeping files you download on an encrypted USB stick. To write your story, you might reboot Tails from a USB stick, with networking disabled but with your persistent storage enabled, for extra security while you prepare your document.

In general, it is best to try to keep anything you want to access in future on strongly encrypted removable media. The default is AES which is becoming worrisome because it so old, and critical features of AES are not yet well understood by the public cryptographic community.

Because different activities and different media engage different Threat Models, it is best to use a small number of different media devices for these different roles, ideally using different strong passphrases, ideally ones you have memorized. EFF's diceware offers a good way to construct long but memorable passphrases with known entropy. Unfortunately, with quantum computing now upon us (or nearly so), you need long passphrases.

Generating long random passphrases requires thought and preparation. Your options include:

o pseudo-random number generator in Tails booted from DVD (less likely that someone who subjects your computer to NSA-quality media recovery will be able to find the random seed you used but you should give the computer enough time to collect good entropy--- if you have trouble generating a long gpg key this is a warning that you probably cannot use this method),

o pseudo-random number generator in your persistent OS (probably better entropy),

o commercial entropy key (but tests suggest these do not produce good entropy, possibly because NSA likes it that way),

o off-line non-computer generation of entropy such as Diceware (but tests suggest cheap dice do not produce good entropy, possibly because NSA likes it that way).

Some people like to use all of these in an entropy smorgasbord.

You can find applications at CRAN and in Debian which generate Diceware passphrases from simulated dice throws. This is a convenient way to get a feel for what kinds of memorization training you will need.

To intelligently generate passphrases you need to start with a threat model such as this:

1 E12 trials per second: Snowden's estimate for NSA supercomputer brute forcing rate c. 2013

1 E4 quantum factor (possibly an underestimate)

3 E7 seconds per year

1 E2 years per century

----

3 E25 alternative passphrases needed

Diceware uses a wordlist with random choices made using pentathrows (five dice):

There are 6^5 = 7776 possible pentathrows, so we compute how many words we need like this:

7776^6 = 2 E23 (good enough for 2013?)

7776^7 = 2 E27 (good enough for 2018?)

But rattling five cubical dice in a box, as Micah Lee recommends, appears to result in like faces sticking together, which very drastically reduces the entropy. Because cheap playing dice are not likely to be adequate for our needs, some people use a computer which they never connect to another device, but use only to generate simulated pentathrows, regularly checking with tools like dieharder the quality of the entropy output by R's default pseudo-random number generator (Mersenne twister, which is much better than the generators which your OS probably uses).

The latest EFF passphrase construction scheme uses "fandom" wordlists (Tolkien loving Tor node operators take note!) with random choices made using trithrows of a dodecahedral die:

There are 20^3 = 8000 possible trithrows, but the wordlists have size 4000, so:

4000^7 = 2 E25 (good enough for 2016?)